Crate sim_lib_compute_cuda 

Optional runtime-loaded CUDA/cuBLAS tensor compute provider.

The CUDA provider has no link-time CUDA dependency. It validates driver, runtime, cuBLAS, and cuBLASLt entry points at runtime and exports a tensor site only while the live library handles remain retained. Dense f32 matmul executes through cuBLAS and retains the result in device storage. Half-family requests fail closed until a cuBLASLt execution path exists.

Structs

ComputeCudaLib

Loadable library that registers a CUDA compute site when runtime validation succeeds.

CudaAbiEvidence

Dynamic-library ABI evidence required by the CUDA provider.

CudaAllocation

Opaque CUDA allocation evidence.

CudaLibrarySet

Loaded CUDA runtime libraries kept alive for function-pointer validity.

CudaLoadError

CUDA dynamic-loading failure.

CudaResidentStorage

Resident storage for a CUDA matmul result.

CudaRuntimeLoader

Real dynamic loader using platform CUDA shared libraries.

CudaRuntimeProbe

Result of CUDA runtime discovery.

CudaSymbolEvidence

One validated runtime symbol.

CudaTensorExecutor

Tensor executor backed by validated CUDA/cuBLAS ABI evidence.

FakeCudaLoader

Fake loader for deterministic tests.

Statics

RECIPES

Cookbook recipes for this lib, embedded at build time.

Traits

DynamicCudaLoader

Loader abstraction used by real and fake CUDA discovery.

Functions

compute_cuda_capability

Capability required before a CUDA hardware tensor site can be realized.

compute_cuda_lib_symbol

Stable symbol for the CUDA runtime library.

compute_cuda_site_symbol

Site symbol exported by a successful CUDA runtime.

cuda_executor_symbol

Stable symbol for the CUDA tensor executor.

discover_cuda_runtime

Discovers CUDA using the real platform dynamic loader.